Detection is the central problem in real-word spelling correction
This addresses a key bottleneck in natural language processing for applications like text editing, though it is incremental in refining the understanding of the problem.
The paper argues that detection, not candidate selection, is the central challenge in real-word spelling correction, showing that while simple trigram models can discriminate intended words from random variations with high accuracy, they fail when every word is a potential error with many candidates.
Real-word spelling correction differs from non-word spelling correction in its aims and its challenges. Here we show that the central problem in real-word spelling correction is detection. Methods from non-word spelling correction, which focus instead on selection among candidate corrections, do not address detection adequately, because detection is either assumed in advance or heavily constrained. As we demonstrate in this paper, merely discriminating between the intended word and a random close variation of it within the context of a sentence is a task that can be performed with high accuracy using straightforward models. Trigram models are sufficient in almost all cases. The difficulty comes when every word in the sentence is a potential error, with a large set of possible candidate corrections. Despite their strengths, trigram models cannot reliably find true errors without introducing many more, at least not when used in the obvious sequential way without added structure. The detection task exposes weakness not visible in the selection task.